Deep learning improves dermatology diagnoses across 2 broad patient populations

Researchers in South Korea have shown that AI trained on photos of both Asian and Caucasian patients can help dermatologists more accurately diagnose many diseases and disorders of the skin in both subpopulations.

In fact, the technique showed it could even help people with zero medical training know when they need to be seen by a skin specialist for a likely skin cancer.

As reported in the Journal of Investigative Dermatology, the team’s convolutional neural network (CNN) identified, classified and recommended treatments for some 134 dermatological diagnoses “with a performance that is comparable to that of the experts.”

To get to that point, the investigators collected 220,000 photos of 174 skin conditions to train their algorithm, then compared its initial performance against that of 21 experienced dermatologists, 26 dermatology residents and 23 members of the general public.

In the algorithm’s first head-to-head, it proved about as able as the residents but not as precise as the seasoned dermatologists.

However, after the initial interpretations, all participants were shown the algorithm’s results—and collectively modified their calls with significant success: The 47 expert dermatologists and residents improved their sensitivity scores from 77.4% to 86.8%. The 23 laypersons saw their sensitivity scores catapult from 47.6% to 87.5%.

“Notably, based on the initial result, half of the malignancies would have been missed by the general public without referral to specialists,” the authors comment in their discussion section.

In a news release sent by journal publisher Elsevier, lead investigator Jung-Im Na, MD, PhD, of Seoul National University suggests the CNN might supply dermatologists with “augmented intelligence.”

In so doing, Na adds, the tool would support rather than replace physicians.

Further, Na says, the algorithm combined with a smartphone for photographing skin lesions “could encourage the public to visit specialists for cancerous lesions such as melanoma that might have been neglected otherwise.”

Na cautions care for members of the general public using the method on their own, as images of insufficient quality or poor composition may affect the algorithm’s accuracy.

Still, if the algorithm's performance can be reproduced in a clinical setting, “it will be promising for the early detection of skin cancer with a smartphone,” Na says before expressing hope for future studies to gauge the utility and performance of the CNN in clinical settings.

The study is available in full for free.